Second VoyageBlog
← All case studies
Manufacturing and mechanical design4 min read

Teaching an AI to drive CAD software, and to check its own work

An MCP bridge that lets Claude design parts in Alibre Design, measuring every step against intent. It proved the agent-harness pattern we now use for CAM.

Alibre Harness · Internal R&D (for Brookline Industries' engineering work)

Engineering software is where small manufacturers lose hours. Drawing a mounting plate, a bracket or a fixture plate is skilled but repetitive work, and the desktop CAD tools that do it were built for a person with a mouse, not for automation. AI models can describe a part in words, but describing a part and producing a correct, dimensioned solid model are different things. The question we wanted to answer: can an AI agent operate real CAD software, and can we trust what it builds?

At a glance

Client Internal R&D, for the engineering work at Brookline Industries
Industry Manufacturing and mechanical design
What we built A bridge (an MCP server) that gives Claude 22 tools to design parts and assemblies inside Alibre Design, a desktop CAD package
Where AI fits Claude is the operator. It plans the part, builds it in small steps, and measures each step. The bridge makes every action checkable.
Stack Python talking to Alibre's Windows automation interface, exposed to Claude through the Model Context Protocol (MCP)
Status An MCP bridge that lets Claude design parts in Alibre Design, checking every step

The old way

  1. An engineer reads a sketch or spec and opens the CAD software.
  2. They draw each feature by hand: sketch, extrude, holes, chamfers.
  3. They check dimensions by eye or with the measure tool.
  4. They export STEP or drawing files for the shop.
  5. A small change to the spec means repeating much of it.

What we built

The harness connects Claude to a running copy of Alibre Design. Claude gets tools for sessions, building geometry (through a Python helper library that wraps every verified API call), measuring (bounding box, volume, mass, inertia), viewport screenshots from standard views, editing parameters, materials, checking assemblies for interference, importing and exporting STEP and other formats, and generating 2D drawings.

The important part is the working method, written into the agent's own instructions:

  1. Build in small steps: one sketch and one feature at a time.
  2. Verify after every step, numerically. Every call returns the part's size and face count, and Claude compares that to the intended dimensions. Screenshots are only for a quick shape check, never for dimensions.
  3. On an error, read the returned feature tree, fix it, and retry, with a limited number of repair rounds before going back to the documentation.
  4. Export at the end.

How the workflow changes

Step Before After
Starting a part Open the CAD tool and sketch Describe the part in plain language: "a 60×40×8 mm plate with four 5 mm corner holes on a 50×30 pattern"
Building features Draw each one by hand Claude builds it feature by feature in the real CAD application, visible on screen as it goes
Checking dimensions By eye, or spot measurements Measured after every step and compared numerically to intent
Fixing mistakes Find it later in the shop Caught at the step that caused it, with the feature tree to diagnose it
Delivering files Manual export STEP, STL and drawing exports as the final step

Inside the build

flowchart LR A[Plain-language<br/>part request] --> B[Claude plans<br/>the next feature] B --> C[Build one feature<br/>in Alibre] C --> D{Measure:<br/>size, volume,<br/>face count} D -- matches intent --> E{Part complete?} D -- mismatch or error --> F[Read feature tree,<br/>repair and retry] F --> C E -- no --> B E -- yes --> G[Export STEP / drawing]

Small steps, each one measured. This is the loop the Fusion 360 harness later formalized into quality gates.

Under the hood

  • A regression suite of 125 known parts and assemblies. Each is built through the same tools Claude uses and checked against expected dimensions, volume and face counts, calculated where possible and locked to known-good values otherwise. 46 come from real design-for-manufacturing test specifications.
  • Assemblies checked numerically. For assemblies, the suite asserts where each component actually ends up, not just that constraints were added, across all ten constraint types, multi-level sub-assemblies and save-and-reopen.
  • Facts, not assumptions. A first proof-of-concept phase verified the vendor API's real behavior (14 of 14 checks passing) and wrote down every quirk. Where a native feature turned out to be broken on this software version, the agent's instructions name the working fallback.
  • Survives software updates. The bridge finds the installed version automatically, and it carried through a vendor service pack mid-development with no changes.

By the numbers

  • 22 tools that give Claude control of Alibre Design.
  • 125 known parts and assemblies in the regression suite, 46 of them from real design-for-manufacturing test specifications.
  • 14 of 14 vendor-API behavior checks passing in the proof-of-concept phase.

The harness proved a pattern: an AI agent can operate desktop engineering software reliably when every action is measured against intent. That pattern carries straight into our Fusion 360 harness, which applies the same discipline to CAM programming and adds formal stop-the-line quality gates.

Example scenario

Example scenario (illustrative): A machine shop makes dozens of fixture plates a year, each a variation on the same theme: a size, a bolt pattern, a few counterbores. An engineer types the spec into a chat. The agent builds the plate in the shop's existing CAD software, measuring each feature as it goes, and hands back a STEP file and a drawing. The engineer reviews a finished, dimension-checked model instead of drawing it from scratch.

What this means for your business

The reusable capability is AI agents for legacy desktop software: connecting a capable model to the tools you already own, with every action checked. It fits:

  • Job shops and fabricators drawing repetitive parts and fixtures
  • Engineering firms with Windows CAD, CAM or analysis tools that lack a modern API
  • Product companies with configurable parts (sizes, patterns, variants)
  • Any team whose expert hours go into driving software step by step

Second Voyage builds AI systems that are held to a standard you can check. Talk to us about yours.

Work with us

Have a business that runs from your head? Start with the AI audit.

Start your AI audit

Email

hello@second.voyage